Upload
others
View
11
Download
0
Embed Size (px)
Citation preview
Bayesian Analysis of ProgressivelyCensored Competing Risks Data
Debasis Kundu ∗& Biswabrata Pradhan †
Abstract
In this paper we consider the Bayesian inference of the unknown parameters ofthe progressively censored competing risks data, when the lifetime distributions areWeibull. It is assumed that the latent cause of failures have independent Weibulldistributions with the common shape parameter, but different scale parameters. Inthis article, it is assumed that the shape parameter has a log-concave prior densityfunction, and for the given shape parameter, the scale parameters have Beta-Dirichletpriors. When the common shape parameter is known, the Bayes estimates of the scaleparameters have closed form expressions, but when the common shape parameter isunknown, the Bayes estimates do not have explicit expressions. In this case we proposeto use MCMC samples to compute the Bayes estimates and highest posterior density(HPD) credible intervals. Monte Carlo simulations are performed to investigate theperformances of the estimators. Two data sets are analyzed for illustration. Finallywe provide a methodology to compare two different censoring schemes and thus findthe optimum Bayesian censoring scheme.
Key Words and Phrases: Latent failure time model; Type-II progressive censoring
scheme; Markov Chain Monte Carlo; Credible interval; Optimum censoring scheme.
∗Department of Mathematics and Statistics, Indian Institute of Technology Kanpur, Pin 208016, In-dia, and a visiting Professor at King Saud University, Saudi Arabia, Corresponding author, e-mail:[email protected]†SQC & OR Unit, Indian Statistical Institute, 203 B.T. Road, Kolkata, Pin 700108, India
1
2
1 Introduction
In many life-testing studies, often the failure of items or individuals may be associated
to more than one cause. These ‘risk factors’ in some sense compete with each other for
the failure of the experimental unit. Due to this reason, in the statistical literature it is
well known as the competing risks model. Several examples can be found, see for example
Crowder [10], where failure may occur due to more than one cause. In analyzing such data
set, the investigator is naturally interested in the assessment of a specific risk in presence of
other risk factors.
In analyzing the data for competing risks model, ideally the data consists of a failure time
and the associated cause of failure. The causes of failure may be assumed to be independent
or dependent. Although the assumption of dependence seems more reasonable, but there is
some concern about the identifiability issue of the competing risks model. Several authors, see
for example Kalbfleish and Prentice [13], Crowder [10], argued that without the information
of covariates, it is not possible to test the assumption of the failure time distributions of the
competing causes, just based on the observed data.
In this paper we use the latent failure time modeling of Cox [9] for analyzing competing
risks data. In the latent failure time modeling, it is assumed that the competing causes of
failures are independently distributed. Here it is further assumed that the lifetime distribu-
tions of the competing causes follow Weibull distributions with the same shape parameter,
but different scale parameters. It may be mentioned that the assumption of the common
shape parameter for the Weibull distribution in case of competing risks of model is not very
unrealistic, see for example Rao et al. [26], Mukherjee and Basu [19], Kundu and Basu [16]
and the references therein.
3
Therefore, if Ti denotes the lifetime of the i-th individual then
Ti = min{Xi1, · · · , XiM},
where Xi1, · · · , XiM are the latent failure times of the M different causes for the i-th individ-
ual. According to the latent failure time model assumption, Xi1, · · · , XiM are independently
distributed. Moreover, Xi1, · · · , XiM are not observable, only Ti is observable and the indica-
tor J such that XiJ = min{Xi1, · · · , XiM} is observable. In this paper it is further assumed
that Xij for j = 1, · · · ,M , follows a Weibull distribution with the probability density function
(PDF)
f(t;α, λj) =
αλje
−λjtαtα−1 if t > 0
0 if t ≤ 0,(1)
here α > 0, λj > 0 are the shape and scale parameters of the Weibull distribution with the
PDF (1) and it will be denoted as WE(α, λj). In this paper, from now on it is assumed
that M = 2 for notational convenience, although all the results presented here are valid for
general M .
Censoring is very common in most of the life-testing and reliability studies, because
quite often the experimenter is unable to obtain complete information on lifetimes of all the
items/individuals. Although, Type-I and Type-II censoring schemes are two most popular
censoring schemes, but recently progressive censoring scheme has received considerable at-
tention in the statistical literature due to its wide scale applicability, see for example Viveros
and Balakrishnan, [27], Balasooriya et al. [6], Ng et al. [21], Kundu [15], Pradhan and
Kundu [22], the review article by Balakrishnan [3] and the references cited therein. Al-
though, extensive work has been done on progressive censoring scheme, not much work has
been done in the competing risk set up.
In this paper we consider the Bayesian analysis of the competing risks data, when the
lifetime distributions are Weibull with the same shape but different scale parameters. For
4
the Bayesian inference of the unknown parameters, we need to assume some priors on the
unknown parameters. If the common shape parameter α is known, the convenient but
quite general conjugate priors on the scale parameters are the Beta-Gamma (for M > 2,
it is Dirichlet-Gamma) priors, see Pena and Gupta [25] in this respect. In this case, the
explicit expressions of the Bayes estimates can be obtained under the squared error loss
function. But when the common shape parameter is unknown, it is known that in this case
the continuous conjugate priors do not exist, see for example Kaminskiy and Krivtsov [14]
in this connection. We use the same conjugate priors on the scale parameters, even when α
was unknown. We have not assumed any specific prior on α, it is simply assumed that the
support of the prior on α is on (0,∞), and that it has a log-concave density function. Note
that the assumption of log-concave density function of the prior distribution is quite common
in Bayesian analysis, see Berger and Sun [5], and many common distribution functions, for
example normal, log-normal, gamma and Weibull may have log-concave density functions.
Based on the above prior distributions, we obtain the joint posterior distribution of
the unknown parameters. As expected the Bayes estimates cannot be obtained in explicit
forms. We propose to use MCMC samples to compute Bayes estimates and approximate
highest posterior density (HPD) credible intervals of the unknown parameters. One can
also apply Lindley’s approximation to compute Bayes estimates. It is observed that our
method can be easily extended even when some of the causes of failures are unknown. We
compare the performances of Bayes estimates and HPD credible intervals with the classical
maximum likelihood estimators (MLEs). It is observed that if we do not have any prior
information, the performances of the MLEs and the Bayes estimators are quite comparable.
But with informative priors, the Bayes estimates behave much better than the MLEs, as
expected. Although we have derived the Bayes estimates of the unknown parameters based
on progressively censored data, the proposed method is easily extendable for other censoring
schemes. In survival analysis, data are mainly random censored. We have analyzed such a
5
data set in Section 4.2 to illustrate our methodology.
The second aim of this manuscript is to provide the methodology to compare two dif-
ferent sampling schemes, and hence in turn to compute the optimal censoring scheme in
presence of competing risks. Finding the optimal progressive censoring scheme is an impor-
tant problem, and it has received considerable attention in the recent statistical literature
due to its practical applicability. In the progressive censoring scheme, an optimal censoring
scheme means, for fixed n and m, the choice of {R1, · · · , Rm}, which provides the maximum
information of the unknown parameters. Unfortunately, not much work has been done in
this direction in presence of competing risks.
Using the idea of Zhang and Meeker [28], we have proposed two information measure
of the unknown parameters for a given progressive censoring scheme, when the competing
causes of failure are present. We have provided the optimal censoring schemes based on
different criteria and compared the results with the traditional Type-II censoring also. It
is observed that the relative efficiencies of the Type-II censoring schemes are quite close to
one.
The rest of the paper is organized as follows. In section 2 we provide the model for-
mulation and prior assumptions. Posterior analysis and Bayesian inference of the unknown
parameters are provided in section 3. Numerical simulation results and the analysis of two
data sets are presented in section 4. In section 5 we provide the optimal censoring plan, and
finally we conclude the paper in section 6.
2 Model Formulation and Prior Assumptions
In this section we introduce the model, and the available data. We also provide the necessary
prior assumptions for further development.
6
2.1 Model Formulation and Available Data
Suppose n identical items are put on a test at the time point zero, with the lifetime of the
n-items are denoted by T1, · · · , Tn. It is assumed that for each i, Ti = min{Xi1, Xi2}, where
Xi1 ∼WE(α, λ1), Xi2 ∼WE(α, λ2) and they are independently distributed. Therefore, Ti ∼
WE(α, λ1 + λ2), and moreover it is assumed that T1, · · · , Tn are independently distributed.
At the time of each failure, the failure time and the corresponding cause of failure is observed.
The integer m < n is pre-fixed, and R1, · · · , Rm are pre-fixed integers such that
R1 + · · ·+Rm = n−m. (2)
At the time of the first failure, say t1, R1 of the remaining units are randomly chosen and
removed. Similarly, at the time of the second failure, say t2, R2 of the remaining units are
chosen and removed, and so on. Finally at the time of the m-th failure time, tm, the rest of
the units, Rm = n−m−R1− · · ·−Rm−1 are removed and the experiment stops. Therefore,
a progressively censored competing risk data will be as follows:
{(t1, δ1, R1), · · · , (tm, δm, Rm)}. (3)
Here δ1, · · · , δm denote the m causes of failures at the time points t1, · · · , tm respectively, and
for each i, δi takes a value either 1 and 2. Therefore, for a given R1, · · · , Rm, we have the
following m observations;
{(ti, 1); i ∈ I1}, and {(ti, 2); i ∈ I2}, (4)
here
I1 = {i; δi = 1}, and I2 = {i; δi = 2}.
7
2.2 Prior Assumptions
When the common shape parameter α is known, the scale parameters have conjugate priors.
Using the idea of Pena and Gupta [25], it is assumed that λ = λ1 + λ2 has a Gamma(a0, b0)
prior, say π0(·| a0, b0·). Here the PDF of Gamma(a0, b0) for λ > 0 is;
π0(λ| a0, b0) =ba00
Γ(a0)λa0−1e−b0λ, (5)
and 0 otherwise. Given λ, λ1/λ has Beta(a1, a2), say π1(·|a1, a2) prior, i.e.
π(λ1/λ| a1, a2) =Γ(a1 + a2)
Γ(a1)Γ(a2)
(λ1
λ
)a1−1 (1− λ1
λ
)a2−1
(6)
for λ1/λ > 0, and 0 otherwise. Here all the hyper-parameters a0 > 0, b0 > 0, a1 > 0, a2 > 0.
It will be shown that when the common shape parameter is known, the above priors are the
conjugate priors. After simple transformation, the joint prior of λ1 and λ2 becomes;
π(λ1, λ2|a0, b0, a1, a2) =Γ(a1 + a2)
Γ(a0)× (b0λ)a0−a1−a2 × ba1
0
Γ(a1)λa1−1
1 e−b0λ1 × ba20
Γ(a2)λa2−1
2 e−b0λ2 .
(7)
This is the Beta-Dirichlet distribution, and it will be denoted by BD(b0, a0, a1, a2). Clearly, in
general λ1 and λ2 will be dependent, but when a0 = a1+a2, they are independent. Therefore,
independent priors can be obtained as a special case of (7). It may be easily observed that
the covariance of λ1 and λ2 can be positive or negative depending on a0 > a1 + a2 or
a0 < a1 + a2. The following result will be useful for further development, whose proof can
be easily obtained from Theorem 2 of Pena and Gupta [25].
Result: If (λ1, λ2) ∼ BD(b0, a0, a1, a2), then for i = 1, 2,
E(λi) =a0ai
b0(a1 + a2)and V (λi) =
a0aib20(a1 + a2)
×{
(ai + 1)(a0 + 1)
a1 + a2 + 1− a0aia1 + a2
}. (8)
When the common shape parameter is known, the above priors are the conjugate priors.
But when the shape parameter is not known, the conjugate priors do not exist. In this case
8
it is assumed that λ1 and λ2 have the same Beta-Dirichlet prior as defined (7) and the prior
on α is independent of (λ1, λ2). No specific form of prior on α has been assumed here. It is
only assumed that the absolute continuous prior π(α) on α has a positive support on (0,∞)
and the PDF of π(α) is log-concave and it is independent of (λ1, λ2). Although, in general
the choice of the hyper-parameters are very important in practice, it is not pursued here.
3 Posterior Analysis and Bayes Inference
In this section, we provide the Bayes estimates of the unknown parameters and the cor-
responding credible intervals, when the common shape parameter is known and when it
is unknown based on the priors assumed in the previous section. We mainly assume the
squared error loss (SEL) function, although any other loss function can be easily considered,
without much of a difficulty.
3.1 Common Shape Parameter Known
Based on the observed sample {(t1, δ1, R1), · · · , (tm, δm, Rm)}, the likelihood function is;
l(data| α, λ1, λ2) ∝ αmλm11 λm2
2 e−(λ1+λ2)∑m
i=1(Ri+1)tαi ×
m∏i=1
tα−1i . (9)
Here m1 and m2 denote the number of elements in the set I1 and I2 respectively. For known
α, when λ1 and λ2 have the joint priors as given in section 2, it can be easily observed that
the joint posterior distribution of λ1 and λ2, i.e.
l(λ1, λ2|data, α) ∝ BD
(b0 +
m∑i=1
(Ri + 1)tαi , a0 +m1 +m2, a1 +m1, a2 +m2
). (10)
Therefore, with respect to the squared error loss function, the Bayes estimates of λ1 and λ2
are
λ1B =(a0 +m1 +m2)(a1 +m1)
(b0 +∑mi=1(Ri + 1)tαi )(a1 + a2 +m1 +m2)
(11)
9
and
λ2B =(a0 +m1 +m2)(a2 +m2)
(b0 +∑mi=1(Ri + 1)tαi )(a1 + a2 +m1 +m2)
. (12)
The corresponding posterior variances are
V (λ1B) = A1 ×B1, and V (λ2B) = A2 ×B2, (13)
respectively, where for j = 1, 2,
Aj =(a0 +m1 +m2)(aj +mj)
(b0 +∑mi=1(Ri + 1)tαi )2(a1 + a2 +m1 +m2)
(14)
Bj =
{(aj +mj + 1)(a0 +m1 +m2 + 1)
(a1 + a2 +m1 +m2 + 1)− (aj +mj)(a0 +m1 +m2)
(a1 + a2 +m1 +m2)
}. (15)
Under the assumptions of non-informative priors, i.e. a0 = b0 = a1 = a2 = 0, the Bayes
estimates of λ1 and λ2 become
λ1B =m1∑m
i=1(Ri + 1)tαi, and λ2B =
m2∑mi=1(Ri + 1)tαi
, (16)
and they can be easily seen to be the MLEs of λ1 and λ2 respectively.
Note that although the Bayes estimates can be obtained in explicit forms, the corre-
sponding highest posterior density (HPD) credible intervals cannot be obtained explicitly.
But it is possible to generate MCMC samples by direct sampling from the joint posterior
density function, and they can be used to construct HPD credible intervals of λ1 and λ2.
The details will be explained later.
3.2 Common Shape Parameter Unknown
In this subsection we consider the important case when the common shape parameter is
unknown, which is most likely to happen in practice. In this case based on the priors on λ1,
λ2 and α, as it has been assumed in the previous section, the joint density function of on λ1,
10
λ2, α and data is
l(data| α, λ1, λ2)π1(λ1, λ2| a0, b0, a1, b1)π2(α). (17)
Based on (7), the joint posterior density of α, λ1 and λ2 is
l(α, λ1, λ2| data) =l(data| α, λ1, λ2)π1(λ1, λ2| a0, b0, a1, b1)π2(α)∫∞
0
∫∞0
∫∞0 l(data| α, λ1, λ2)π1(λ1, λ2| a0, b0, a1, b1)π2(α)dαdλ1dλ2
. (18)
Therefore, the Bayes estimate of any function of α, λ1 and λ2, say g(α, λ1, λ2) under the
squared error loss function can be obtained as
gB(α, λ1, λ2) = Eα,λ1,λ2|data(g(α, λ1, λ2))
=
∫∞0
∫∞0
∫∞0 g(α, λ1, λ2)l(data| α, λ1, λ2)π1(λ1, λ2| a0, b0, a1, b1)π2(α)dαdλ1dλ2∫∞
0
∫∞0
∫∞0 l(data| α, λ1, λ2)π1(λ1, λ2| a0, b0, a1, b1)π2(α)dαdλ1dλ2
.
(19)
Note that in most of the situation, it is not possible to compute (19) explicitly. There are
several methods available to approximate (19), but they do not provide the HPD credible
intervals. We propose to use MCMC samples generated by direct sampling method to
approximate (19) and also to compute the HPD credible intervals of the unknown parameters.
The following results will be used for generating samples.
Theorem 1: The joint conditional distribution of λ1 and λ2 given α is
BD
(b0 +
m∑i=1
(Ri + 1)tαi , a0 +m1 +m2, a1 +m1, a2 +m2
). (20)
Proof: It is trivial and it has already been mentioned in the previous subsection.
Theorem 2: The conditional density of α given the data is log-concave.
Proof: Observe that
l(α| data) ∝ π2(α)αmm∏i=1
tα−1i × 1
(b0 +∑mi=1(Ri + 1)tαi )a0+m1+m2
. (21)
Now the result follows using Theorem 2 of Kundu [15].
11
Devroye [11] has suggested to generate samples from a general log-concave density func-
tion. Therefore, using Theorem 1 and Theorem 2, it is possible to generate MCMC samples
from the joint posterior density function (18). We use the following algorithm:
Algorithm:
Step 1: Generate α from (21) using the method suggested by Devroye [11].
Step 2: For given α, generate (λ1, λ2) from (20), as suggested in Appendix A.
Step 3: Repeat Step 1 and Step 2, M times, and obtain (α1, λ11, λ21), · · · , (αM , λ1M , λ2M).
Step 4: The Bayes estimates α, λ1 and λ2 with respect to SEL function respectively as
αB =1
M
M∑k=1
αk, λ1B =1
M
M∑k=1
λ1k and λ2B =1
M
M∑k=1
λ2k.
Step 5: The corresponding posterior variances can be obtained respectively as
1
M
M∑k=1
(αk − αB)2 ,1
M
M∑k=1
(λ1k − λ1B
)2and
1
M
M∑k=1
(λ2k − λ2B
)2.
Step 6: To obtain credible interval of α, we order {αi}, as α(1) < α(2) < · · · < α(M). Then
100(1-2β)% credible intervals of α become
(α(j), α(j+M−2Mβ)), for j = 1, · · · , 2Mβ.
Therefore, 100(1-2β)% HPD credible interval of α becomes (α(j∗), α(j∗+M−2Mβ)), where
j∗ is such that
α(j∗+M−2Mβ) − α(j∗) ≤ α(j+M−2Mβ) − α(j)
for all j = 1, · · · , 2Mβ. Similarly, we obtain the HPD credible intervals for λ1 and λ2.
12
4 Numerical Results and Data Analysis
In this section we conduct a simulation study to investigate the performance of the proposed
Bayes estimators and analyze a data set for illustration. The simulation study is carried out
when both the shape and scale parameters are unknown.
4.1 Simulation study
We use different sample sizes n, different effective sample sizes m and the following sets
of parameters {α = 1, λ1 = 0.6, λ2 = 0.4} and {α = 2, λ1 = 0.6, λ2 = 0.4}. For the
simulation study, it is assumed that α has gamma prior with the shape parameter a and
scale parameter b. We consider non-informative and informative priors both for the shape and
scale parameters. In case of non-informative prior we take a = b = a0 = b0 = a1 = a2 = 0.
We call it as prior 0. The informative priors for the two sets of parameter values are given
in Table 1. It may be noted that prior 1 and prior 3 are selected in such way that prior
means are same as the original means. Hence prior 1 and prior 3 may be considered as the
calibrated priors. Whereas prior 2 and prior 4 are not calibrated and they have been selected
arbitrarily. Performance of the estimators are studied in terms of their biases and the mean
squared errors (MSEs). We also compute the average lengths and the coverage percentages
of the 95% HPD credible intervals based on MCMC samples generated by direct sampling.
Table 1: Different informative priors for the two sets of parameters
Parameters Priorα = 1, λ1 = 0.6, λ2 = 0.4 Prior 1: a=b=5, a0=b0=2, a1 = 0.6 and a2 = 0.4
Prior 2: a= 2.5, b=1, a0= 1, b0=1.5, a1 = 0.8 and a2 = 0.9α = 2, λ1 = 0.6, λ2 = 0.4 Prior 3: a= 5, b=2.5, a0=b0=2, a1 = 0.6 and a2 = 0.4
Prior 4: a= 5, b=1.5, a0= 2.5 b0= 1, a1 = 1.5 and a2 = 0.6
For both the parameter sets, we considered the following sampling schemes:
13
• Scheme 1 (CS-1): n = 15, m = 12, R1 = · · · = R11 = 0, R12 = 3;
• Scheme 2 (CS-2): n = 30, m = 15, R1 = · · · = R14 = 0, R15 = 15;
• Scheme 3 (CS-3): n = 40, m = 15, R1 = · · · = R14 = 0, R15 = 25;
• Scheme 4 (CS-4): n = 40, m = 30, R1 = · · · = R29 = 0, R30 = 10;
• Scheme 5 (CS-5): n = 40, m = 30, R2 = · · · = R30 = 0, R1 = 10;
• Scheme 6 (CS-6): n = 40, m = 30, R2 = · · · = R29 = 0, R1 = R30 = 5;
• Scheme 7 (CS-7): n = 40, m = 30, R1 = · · · = R14 = R17 = · · · = R30 = 0, R15 =
R16 = 5.
Note that the sampling schemes CS-1, CS-2, CS-3 and CS-4 are the usual Type-II censoring
schemes, that is n − m items are removed at the time of the m-th failure. The sampling
scheme CS-5 is just the opposite of the Type-II censoring scheme and in this case n − m
items are removed at the time of the first failure itself. This particular progressive censoring
scheme is known as first step-censoring and this is a particular case of one-step censoring
introduced by Balakrishnan et al. [4]. It is well known that for fixed n and m, the ex-
pected experimental time of the Type-II censoring scheme is less than the corresponding
first step-censoring scheme. In fact for fixed n and m, the expected experimental time of any
other censoring schemes are always between these two extremes. For example, the expected
experimental time of the sampling scheme CS-6 and CS-7 will be between the schemes CS-4
and CS-5.
For generating the progressively censored Weibull samples, we use the algorithm suggested-
in Balakrishnan and Sandhu [1]. For each data point we assigned the cause of failure as 1 or 2
with probability λ1/(λ1 +λ2) and λ2/(λ1 +λ2), respectively. In each case we calculate Bayes
estimates using 10000 MCMC samples and also the 95% HPD credible intervals. We also
14
compute maximum likelihood estimate (MLE) and 95% confidence interval for the purpose of
comparison with the Bayes estimates. We replicate the process for 1000 times and compute
average estimates, mean squared errors, average length of HPD credible intervals/confidence
interval and coverage percentage. The results are reported in Tables 2-5.
Now we compare the performance of the Bayes estimators under different scenarios and
MLEs in terms of biases, MSEs and length of credible intervals. From the simulation results,
it can be seen that the performance of the Bayes estimates with respect to non-informative
priors and MLEs are quite close, as expected. So, if we have no prior information on the
unknown parameters, it is better to use MLEs than Bayes estimators, because the Bayes
estimators are computationally more expensive. The performance of the Bayes estimators
under informative calibrated priors (prior 1 and prior 3) is better than the MLEs or the
Bayes estimators under non-informative prior and informative uncalibrated priors (prior 2
and prior 4), as expected.
4.2 Data Analysis
In this section, we analyze two data sets for illustrative purpose.
Example 1: We consider the competing risks data set of Nelson [20], which consist of
failure or censoring times for 139 appliances (36 in Group I, 51 in Group II, and 52 in Group
III) subjected to a manual lifetime test. This example has been used by several authors for
illustration, see for example, Crowder [10] and Park and Kulasekera [24]. Although there
are three groups, we consider group II for illustration. There are 51 appliances in Group II.
Fifteen failures are due to mode 11, 6 failures are due to other modes and remaining are
censored observations. For illustrative purpose, we consider failure mode 11 as cause 1 and
other failure failure modes are considered as cause 2. So there are 15 failures due to cause
1 and 6 failures due to cause 2. We first analyze the data assuming that the latent cause
15
of failures have independent Weibull distributions with the different shape (α1 and α2) and
scale parameters (λ1 and λ2). The maximum likelihood estimate of the shape parameters are
α1 = 1.433398 and α2 = 0.983995 with standard errors 0.29338 and 0.35120, respectively.
When we perform the following testing of hypothesis problem: H0 : α1 = α2, we cannot
reject the null hypothesis, this justifies that shape parameters can be taken as equal.
Next we analyze the data by generating progressively type-II censored data from the
original data set. We generate progressively type-II censored sample with m = 12 and
censoring schemes R1 = 5, R2 = 2, R3 = 2, R4 = 2, R5 = 14, R6 = 0, R7 = 0, R8 = 0,
R9 = 3, R10 = 0, R11 = 6 and R12 = 5. The progressively type-II sample is (45, 2), (47,
1), (73, 1), (145, 1), (281, 2), (311, 1), (471, 2), (490, 1), (569, 2), (575, 1), (630, 1), (838,
1). Here we have m1 = 8 failure due to cause 1 and m2 = 4 failures due to cause 2. The
maximum likelihood estimates of α, λ1 and λ2 with standard error in parentheses are 1.34094
(0.31988), 0.000051 (0.00010) and 0.000025 (0.000052), respectively. The asymptotic 95%
confidence intervals of α, λ1 and λ2 based on observed information matrix are (0.71397,
1.96790), (-0.00015, 0.00028), (-0.00008, 0.000128).
Next, we compute the Bayes estimate of α, λ1 and λ2. We compute Bayes estimate under
non-informative priors, since we have no prior information about the unknown parameters.
Based on data, the posterior density function of α can be approximated by the gamma
density function by equating the first two moments. The scale and shape parameters of
the approximate gamma distribution are 17.3962 and 13.0559, respectively. Based on 10,000
MCMC samples, the Bayes estimates of α, λ1 and λ2 are 1.33406 (0.32239), 0.00025 (0.00061)
and 0.00012 (0.00030). The 95% HPD credible intervals are (0.74162, 1.97743), (7.587188e-
010, 0.00105) and (3.793594e-010, 0.00052), respectively.
Example 2: Here we consider another set of data to illustrate our methodology. We
analyze the data of a lung cancer clinical trial being conducted by the Eastern Cooperative
Oncology Group, Lagakos [18]. There are 194 cases with 83 deaths from cause 1 (local
16
spread of disease) and 44 from cause 2 (metastatic spread), the remaining 67 times being
right-censored. We analyze this data set how the proposed method can be applied for random
censored data. The present data set is an example of competing risks data with exactly two
causes of failure. The latent failure times independently Weibull distributed with same shape
parameters but different scale parameters (See Crowder [10]). The data set was analyzed
with different shape parameters to check whether the shape parameters are equal or not. Let
α1 and α2 be the shape parameters of two latent time distributions. The maximum likelihood
estimate of α1 and α2 are α1= 1.30374 and α2= 1.36784 with standard error 0.05731 and
0.08895, respectively. It is clear that the there is no significant difference between the shape
parameter values. In this case also the null hypothesis H0 : α1 = α2 cannot be rejected.
Hence the shape parameter for both the Weibull distribution can be taken as equal. The
justification for Weibull distribution is given in Crowder [10].
The maximum likelihood estimates of α, λ1 and λ2 with standard errors in parentheses
are 1.32547 (0.08547), 0.00767 (0.00240) and 0.00407 (0.00134), respectively. The asymptotic
95% confidence intervals of α, λ1 and λ2 based on observed information matrix are (1.15795,
1.49299), (0.00297, 0.01237) and (0.00144, 0.0067). Next we compute Bayes estimates of α,
λ1 and λ2 under non-informative priors. Based on data, the posterior density function of
α can be approximated by the gamma density function by equating the first two moments.
The scale and shape parameters of the approximate gamma distribution are 240.0418 and
181.3844, respectively. Based on 10,000 MCMC samples, the Bayes estimates of α, λ1 and
λ2 with standard errors in parentheses are 1.32493 (0.08584), 0.00800 (0.00236) and 0.00424
(0.00125). The 95% HPD credible intervals are (1.16285, 1.49920), (0.00397, 0.01277) and
(0.00210, 0.00677).
17
5 Optimal Censoring Scheme
Till this point we have discussed the Bayesian inference of the unknown parameters for pro-
gressively censored competing risks data. It is assumed so far that the progressive censoring
is pre-fixed, i.e. n, m and R1, · · · , Rm are known apriori. But in practice the natural question
is, whether we should choose the progressive censoring scheme based on convenience or based
on some scientific criterion.
In the last few years finding the “optimum” censoring scheme has received considerable
attention in the recent statistical literature, see for example Balakrishnan and Aggarwala
[2] (Chapter 10), Ng et al. [21], Balasooriya et al. [6], Burkschat [7], Burkschat et al. [8]
and the references cited therein. Optimum censoring scheme means, among all the possible
progressive censoring schemes, find the scheme which is “best”, with respect to a certain
criterion. Naturally we would like to choose that censoring scheme which provides maximum
“information” of the unknown parameters. Intuitively, it is quite clear that if we do not have
any restriction on n and m, we should choose n = m and n should be as large as possible.
In practice often we do not have any choice of n and m, and hence from now onwards if
nothing is mentioned it is assumed that n and m are fixed.
Now all possible progressive censoring schemes, for fixed n and m, means all possible
choices of R1, · · · , Rm, such R1 + · · · + Rm = n −m. Therefore, in this section, if nothing
is mentioned it is assumed that n and m are fixed. There are two issues involved in dealing
with the optimum censoring scheme, namely (i) find a proper criterion, (ii) with respect to
that criterion find the best censoring scheme. Interestingly both are quite important and
none of them is an trivial issue in this case.
18
5.1 Precision Criterion
To find a proper criterion, one needs to define an “information measure” of a given sampling
scheme R = {R1, · · · , Rm}. Once the information measure of a given sampling scheme is
properly defined, it is possible to compare two different sampling schemes, and hence the
“optimum” censoring scheme can be obtained. It is clear that “optimum” censoring scheme
will depend on the “information” measure defined for a given scheme R.
First let us discuss how the “information” measure can be defined for a given R, and how
it can be used to compare two different censoring schemes. In this respect, the Fisher infor-
mation of the given censoring scheme seems to be a natural choice. If only one parameter
is unknown, then comparison between the two Fisher information of two censoring schemes,
simply boils down to compare two real numbers. But if more than one parameter is un-
known, then comparison of two censoring schemes means to compare two Fisher information
matrices, which may not be a trivial issue. Different methods have been proposed in the
literature to compare two Fisher information matrices, for example the trace or determinant
of the Fisher information matrix. But unfortunately it is known that they are not scale
invariant.
Alternatively, the variances of the 100p-th quantile estimator can be used as an infor-
mation measure for a given censoring scheme, and hence can be used quite effectively to
compare two different censoring schemes. Zhang and Meeker [28] and Ng et al. [21] used
this criterion to find optimum censoring schemes in two different problems. It may be men-
tioned that the above criterion is scale invariant, but it may depend on p, and it may be
chosen based on some practical consideration. Zhang and Meeker [28] and Ng et al. [21]
used p = 0.95 and 0.99, but other values of p also can be chosen.
Recently, Pareek et al. [23] proposed the following two information measures for a given
19
progressive censoring scheme R, in the frequentist context and when the competing risks
data are observed. Suppose the p-quantile points of the two latent lifetime distributions are
Tp,1 =[− 1
λ1
ln(1− p)] 1α
, and Tp,2 =[− 1
λ2
ln(1− p)] 1α
(22)
for cause 1 and cause 2 respectively. Then for fixed 0 ≤ w ≤ 1, the Criterion 1 is defined as
C1(R) = wV ar(ln Tp,1
)+ (1− w)V ar
(ln Tp,2
)(23)
here V ar(ln Tp,1
)and V ar
(ln Tp,2
)are the asymptotic variance of the MLEs of ln Tp,1 and
ln Tp,2 respectively, and although they might depend on the censoring scheme R, but we
are not making it explicit for brevity. The weight w should be chosen depending on the
importance the practitioner wants to put on cause 1 or cause 2.
Similarly, Criterion 2 has been proposed as
C2(R) = w∫ 1
0V ar
(ln Tp,1
)dW1(p) + (1− w)
∫ 1
0V ar
(ln Tp,2
)dW2(p), (24)
here w, V ar(ln Tp,1
)and V ar
(ln Tp,2
)are same as defined before. Moreover, the weight
functions W1(p) ≥ 0, W2(p) ≥ 0 and they satisfy∫ 1
0dW1(p) =
∫ 1
0dW2(p) = 1. (25)
The two weight functions W1(·) and W2(·) have to be decided before hand depending on the
problem, see for example Pareek et al. [23] for discussions.
Following the procedure of Kundu [15], the natural modifications of C1(R) and C2(R)
will be
Criterion 1: Let for any censoring scheme R
CB1 (R) = wEdata [Vposterior (lnTp,1)] + (1− w)Edata [Vposterior (lnTp,2)] . (26)
Here Vposterior(ln Tp,1
)and Vposterior
(ln Tp,2
)denote the posterior variance of lnTp,1 and lnTp,2
respectively. Moreover, Edata(·) means unconditional expectation with respect to the data.
20
If we have two censoring schemes say R1 = {R11, · · · , R1
m} and R2 = {R21, · · · , R2
m}, we say
R1 (R2) is better than R2 (R1) if CB1 (R1) < (>) CB
1 (R2).
Criterion 2: If for any censoring scheme R
CB2 (R) = wEdata
[∫ 1
0Vposterior (lnTp,1) dW1(p)
]+ (1−w)Edata
[∫ 1
0Vposterior (lnTp,2) dW2(p)
],
(27)
then we say R1 (R2) is better than R2 (R1) if CB2 (R1) < (>) CB
2 (R2).
It is clear that the computation of (26) and (27) are not very easy. We have used Monte
Carlo simulation technique to approximate (26) and (27), as proposed in Kundu [15]. For
illustrative purposes, we present the optimal sampling scheme for different objective functions
(OF) for selected combination of m and n. Here it is assumed a1 = a2 = b1 = b2 = 1.0,
and a0 = b0 = 2. We have used the minimum trace criterion (OF-1), minimum determinant
criterion (OF-2), and the minimum sum of the variances of the p-th percentile estimators
for different choices of p, namely p = 0.1 (OF-3), p = 0.99 (OF-4). Finally we have also
computed the optimum scheme based on the minimum sum of the weighted variances (OF-5)
under the assumption that W1(p) = W2(p) = 1 for all 0 ≤ p ≤ 1. Moreover in calculating
OF-3, OF-4 and OF-5, it is assumed that w = 1/2.
Note that in all cases the optimization has to be performed numerically. They are discrete
optimization problem. For a given n and m, the optimum censoring scheme with respect to
to a given criterion can be found by exhaustive search for all possible values of Ri’s satisfying
(2). The results are reported in Tables 6 and 7. Since Type-II censoring scheme is a particular
choice of the general progressive censoring scheme, and it is one of the most used censoring
scheme, we report the relative efficiency and also the relative expected experimental time
of the Type-II censoring scheme with respect to the optimal censoring scheme. To compute
relative efficiency, we need expected Fisher information matrix. The expression of expected
Fisher information matrix is given in Appendix B. It is clear that the relative expected
21
experimental time of the Type-II censoring scheme is significantly smaller than the optimal
censoring scheme, but the relative efficiencies are quite high. That justifies the overwhelming
popularity of the Type-II censoring scheme.
6 Concluding remarks
In this work, we have considered the Bayesian analysis of the competing risks data when
they are Type-II progressively censored. Latent failure time modeling of Cox [9] has been
assumed in analyzing the competing risks data. The latent lifetime distributions are assumed
to be Weibull with the same shape but different scale parameters. Although, in this paper
we have mainly dealt with two causes of failure only, the work can be extended to more than
two causes of failure. Prior selection has been suggested using the idea of Pena and Gupta
[25], which is a very flexible prior.
We have proposed different criteria for selecting optimal censoring schemes and presented
some optimal censoring scheme for selected choice of n and m. We further compute the
relative efficiency of the Type-II censoring schemes with the optimal censoring schemes. It
is observed that the relative efficiency of the Type-II censoring schemes is quite high. It
justifies the usefulness of the Type-II censoring scheme in practice.
Acknowledgements:
The auhtors would like to thank one anonymous referee for many constructive suggestions.
22
Appendix A: Generation from Beta-Dirichlet Distri-
bution
In this appendix we will mention how to generate samples from (20). It has already been
shown that
(λ1, λ2) ∼ BD
(b0 +
m∑i=1
(Ri + 1)tαi , a0 +m1 +m2, a1 +m1, a2 +m2
)⇔
(λ1 + λ2) ∼ Gamma
(a0 +m1 +m2, b0 +
m∑i=1
(Ri + 1)tαi
)and
λ1|(λ1 + λ2) ∼ Beta(a1 +m1, a2 +m2).
Therefore, first we generate from a gamma distribution with the shape and scale parameters
as a0 + m1 + m2 and , b0 +m∑i=1
(Ri + 1)tαi respectively, and then we generate from a Beta
distribution with the parameters as a1+m1 and a2+m2 respectively. Note that the generation
from a Beta distribution can be easily performed using the acceptance-rejection principle,
where the dominating function can be taken as the uniform (0, 1).
Appendix B: Expected Fisher Information Matrix
For completeness purposes we have provided the expected Fisher information matrix. Let
us use the following notation;
rj = m− j + 1 +m∑i=j
Ri, cj−1 =j∏i=1
ri, a1,1 = 1, ai,j =j∏
k=1,k 6=i
1
rk − ri.
If the 3× 3 expected Fisher information matrix is denoted by E, then
E =
E11 E12 E13
E21 E22 E23
E31 E32 E33
= −E
∂2 ln l(α,λ1,λ2)
∂α2∂2 ln l(α,λ1,λ2)
∂α∂λ1
∂2 ln l(α,λ1,λ2)∂α∂λ2
∂2 ln l(α,λ1,λ2)∂λ1∂α
∂2 ln l(α,λ1,λ2)∂λ2
1
∂2 ln l(α,λ1,λ2)∂λ1∂λ2
∂2 ln l(α,λ1,λ2)∂λ2∂α
∂2 ln l(α,λ1,λ2)∂λ2∂λ1
∂2 ln l(α,λ1,λ2)∂λ2
2
, (28)
where
E11 =m
α2+
(λ1 + λ2)
α2
m∑j=1
(Rj + 1)h1j
23
E22 =1
λ1(λ1 + λ2), E33 =
1
λ2(λ1 + λ2)
E23 = E32 = 0
E12 = E21 = E13 = E31 =1
α
m∑j=1
(Rj + 1)h2j,
and
h1j =cj−1
(λ1 + λ2)
j∑i=1
ai,jr2i
[π2
6− 2γ + γ2 − 2(1− γ) ln{ri(λ1 + λ2)}+ (ln{ri(λ1 + λ2)})2
]
h2j =cj−1
(λ1 + λ2)
j∑i=1
ai,jr2i
[1− γ − ln(ri(λ1 + λ2))] .
References
[1] Balakrishnan, N. and Sandhu, R.A. (1995), “A simple algorithm for generating
progressively Type-II generated samples”, American Statistician, vol. 49, 229 -
230.
[2] Balakrishnan, N. and Aggarwala, R. (2000), Progressive Censoring: Theory, Meth-
ods and Applications, Birkhauser, Boston.
[3] Balakrishnan, N. (2007), “Progressive censoring methodology: an appraisal”, (with
discussions), Test, vol. 16, 211 - 289.
[4] Balakrishnan, N., Burkschat, M., Cramer, E. and Hoffman, G. (2008), “Fisher
informa- tion based progressive censoring plans”, Computational Statistics and
Data Analysis, vol. 53, 366 - 380.
[5] Berger, J. O. and Sun, D. (1993), “Bayesian analysis for the Poly-Weibull distri-
bution”, Journal of the American Statistical Association, vol. 88, 1412 - 1418.
24
[6] Balasooriya, U. Saw, S.L.C., and Gadag, V.G. (2000), “Progressively censored
reliability sampling plan for Weibull distribution”, Technometrics, vol. 42 160 -
168.
[7] Burkschat, M. (2008), “On optimality of extremal schemes in progressive type-II
censoring”, Journal of Statistical Planning and Inference, vol. 138, 1647 - 1659.
[8] Burkschat, M., Cramer, E. and Kemps, U. (2007), “On optimal schemes in pro-
gressive censoring”, Statistics and Probability Letters, vol. 76, 1032 - 1036.
[9] Cox, D.R. (1959), “The analysis of exponentially distributed lifetimes with two
types of failures”, Journal of the Royal Statistical Society, Ser. B, vol. 21, 411 -
421.
[10] Crowder, M. (2001), Classical Competing Risks Model, Chapman & Hall/ CRC,
New York.
[11] Devroye, L. (1984), “A simple algorithm for generating random variables using
log-concave density”, Computing, vol. 33, 247 - 257.
[12] Hoel, D. G. (1972), “A representation of mortality data by competing risks”, Bio-
metrics, vol. 28, 475-488.
[13] Kalbfleish, J.D. and Prentice, R.L. (1980), The Statistical Analysis of Failure Data,
New York, Wiley.
[14] Kaminskiy, M. P. and Krivtsov, V. V. (2005), “A simple procedure for Bayesian
estimation of the Weibull distribution”, IEEE Transactions on Reliability, vol. 54,
612 - 616.
[15] Kundu, D. (2008), “Bayesian inference and life testing plan for the Weibull distri-
bution in presence of progressive censoring”, Technometrics, vol. 50, 144–154.
25
[16] Kundu, D. and Basu, S. (2000), “Analysis of incomplete data in presence of com-
peting risks”, Journal of Statistical Planning and Inference, vol. 87, 221 - 239.
[17] Kundu, D., Kannan, N. and Balakrishnan, N. (2004), “Analysis of Progressively
Censored Competing Risks Data”, Handbook of Statistics on Survival Analysis ,
Editors: C.R. Rao and N. Balakrishnan, Elsevier Publications, vol. 23, 331-348.
[18] Lagakos, S. W. (1978), “A covariate model for partially censored data subject to
competing causes of failure“, Applied Statistics, vol. 27, 235-241.
[19] Mukhopadhyay, C. and Basu, A. P. (1997), “ Bayesian analysis of incomplete time
and cause of failure data”, Journal of Statistical Planning and Inference, vol. 59,
79-100.
[20] Nelson, W. (1970), “Hazard plotting methods for analysis of life data with different
failure modes“, Journal of Quality Technology, vol. 2, 126-149.
[21] Ng, T., Chan, C.S. and Balakrishnan, N. (2004), “Optimal progressive censoring
plans for the Weibull distribution”, Technometrics, vol. 46, 470 - 481.
[22] Pradhan, B. and Kundu, D. (2009), “On progressively censored generalized expo-
nential distribution”, Test, vol. 18, 497 - 515.
[23] Pareek, B., Kundu, D. and Kumar, S. (2009), “On progressive censored competing
risks data for Weibull distributions”, Computational Statistics and Data Analysis,
vol. 53, 4083 - 4094.
[24] Park, C. and Kulasekera, K. B. (2004), “Parametric inference of incomplete data
with competing risks among several groups“, IEEE Transactions on Reliability, vol.
53, 11-21.
26
[25] Pena, E. A. and Gupta, A. K. (1990), “Bayes estimation for the Marshall-Olkin
exponential distribution”, Journal of the Royal Statistical Society, Ser. B, vol. 52,
379 - 389.
[26] Rao, B.R., Talwalker, S. and Kundu, D. (1991), “Confidence intervals for the rela-
tive risk ratio parameters from survival data under a random epidemiology study”,
Biometrical Journal, vol. 33, 959 - 984.
[27] Viveros, R. and Balakrishnan, N. (1994), “Interval estimation of parameters of life
from progressively censored data”, Technometrics, vol. 36, 84 - 91.
[28] Zhang, Y. and Meeker, W.Q. (2005), “Bayesian life test planning for the Weibull
distribution with the given shape parameter”, Metrika, vol. 61, 237–249.
27
Table 2: The average values of the MLEs and Bayes estimates under different priors alongwith the MSE’s in parentheses when α = 1, λ1 = 0.6 and λ2 = 0.4.
Estimator CS-1 CS-2 CS-3 CS-4 CS-5 CS-6 CS-7MLE α 1.1398 1.1400 1.1455 1.0598 1.0476 1.0534 1.0450
(0.1341) (0.1079) (0.1144) (0.0347) (0.0252) (0.0304) (0.0235)λ1 0.6585 0.7289 0.7952 0.6227 0.6227 0.6172 0.6293
(0.0965) (0.1812) (0.3743) (0.0249) (0.0256) (0.0235) (0.0281)λ2 0.4491 0.4919 0.5372 0.4243 0.4160 0.4204 0.4211
(0.0663) (0.1030) (0.2074) (0.0197) (0.0175) (0.0185) (0.0185)Bayes α 1.1621 1.0986 1.1308 1.0556 1.0412 1.0455 1.0417Prior 0 (0.1431) (0.0919) (0.1053) (0.0339) (0.0255) (0.0307) (0.0273)
λ1 0.66735 0.6658 0.7893 0.6227 0.6133 0.6161 0.6102(0.1281) (0.0980) (0.2062) (0.0228) (0.0212) (0.0236) (0.0253)
λ2 0.4568 0.4604 0.5318 0.4132 0.4127 0.4135 0.4099(0.0625) (0.0422) (0.2442) (0.0163) (0.0164) (0.0147) (0.0154)
Bayes α 1.0778 1.0630 1.0587 1.0469 1.0359 1.0423 1.0300Prior 1 (0.0476) (0.0444) (0.0377) (0.0263) (0.0193) (0.0244) (0.0200)
λ1 0.6420 0.6541 0.6789 0.6232 0.6166 0.6139 0.6213(0.0557) (0.0484) (0.0571) (0.0221) (0.0207) (0.0193) (0.0222)
λ2 0.4329 0.4358 0.4475 0.4103 0.4120 0.4119 0.4177(0.0364) (0.0335) (0.0369) (0.0141) (0.0141) (0.0135) (0.0142)
Bayes α 1.1960 1.1627 1.1415 1.0814 1.0665 1.0924 1.0676Prior 2 (0.1231) (0.0944) (0.0768) (0.0408) (0.0333) (0.0403) (0.0330)
λ1 0.6155 0.6573 0.6703 0.6098 0.5937 0.5981 0.5995(0.0580) (0.0640) (0.0644) (0.0237) (0.0211) (0.0207) (0.0205)
λ2 0.4498 0.4775 0.4817 0.4120 0.4057 0.4079 0.4194(0.0452) (0.0445) (0.0463) (0.0143) (0.0136) (0.0133) (0.0150)
28
Table 3: The average confidence length/HPD credible length and coverage percentage inparentheses when α = 1, λ1 = 0.6 and λ2 = 0.4
Estimator CS-1 CS-2 CS-3 CS-4 CS-5 CS-6 CS-7MLE α 1.1146 1.0691 1.0984 0.6631 0.5807 0.6268 0.5735
(0.96) (0.96) (0.96) (0.96) (0.95) (0.95) (0.96)λ1 0.9906 1.1384 1.4709 0.5832 0.5893 0.5769 0.5855
(0.94) (0.95) (0.96) (0.94) (0.94) (0.94) (0.94)λ2 0.8038 0.8866 1.1093 0.4783 0.4762 0.4730 0.4759
(0.92) (0.95) (0.95) (0.93) (0.93) (0.93) (0.93)Bayes α 1.1161 0.9982 1.0579 0.6549 0.5636 0.6148 0.5570Prior 0 (0.94) (0.96) (0.96) (0.96) (0.94) (0.95) (0.94)
λ1 0.9620 0.9576 1.3094 0.5668 0.5688 0.5623 0.5568(0.93) (0.93) (0.97) (0.95) (0.95) (0.94) (0.93)
λ2 0.7579 0.7401 0.9942 0.4546 0.4572 0.4554 0.4493(0.94) (0.95) (0.95) (0.94) (0.94) (0.93) (0.94)
Bayes α 0.9051 0.8468 0.8337 0.6118 0.5347 0.5836 0.5255Prior 1 (0.98) (0.97) (0.97) (0.96) (0.96) (0.96) (0.95)
λ1 0.8520 0.8365 0.6789 0.5517 0.5545 0.5449 0.5518(0.95) (0.96) (0.97) (0.95) (0.95) (0.95) (0.94)
λ2 0.6805 0.6550 0.4475 0.4421 0.4464 0.4411 0.4476(0.93) (0.94) (0.95) (0.94) (0.94) (0.94) (0.95)
Bayes α 1.0645 0.9778 0.9499 0.6598 0.5658 0.6273 0.5625Prior 2 (0.95) (0.95) (0.97) (0.90) (0.88) (0.89) (0.86)
λ1 0.8435 0.8839 0.9641 0.5464 0.5435 0.5364 0.5408(0.91) (0.95) (0.96) (0.93) (0.93) (0.94) (0.93)
λ2 0.6999 0.7201 0.7681 0.4419 0.4400 0.4365 0.4455(0.93) (0.96) (0.97) (0.94) (0.94) (0.94) (0.94)
29
Table 4: The average values of MLEs and the Bayes estimates under different priors alongwith the MSE’s in parentheses when α = 2, λ1 = 0.6 and λ2 = 0.4.
Estimator CS-1 CS-2 CS-3 CS-4 CS-5 CS-6 CS-7MLE α 2.2799 2.2801 2.2910 2.1195 2.0923 2.1073 2.0876
(0.5364) (0.4314) (0.4575) (0.1388) (0.0988) (0.1214) (0.0928)λ1 0.6585 0.7289 0.7952 0.6227 0.6230 0.6172 0.6296
(0.0965) (0.1811) (0.3743) (0.0249) (0.0257) (0.0235) (0.0278)λ2 0.4491 0.4919 0.5372 0.4243 0.4177 0.4204 0.4207
(0.0663) (0.1030) (0.2074) (0.0972) (0.0179) (0.0185) (0.0182)Bayes α 2.2754 2.2645 2.2465 2.1112 2.0824 2.0910 2.0835Prior 0 (0.4719) (0.4050) (0.4186) (0.1355) (0.1021) (0.1226) (0.1090)
λ1 0.6763 0.7351 0.8126 0.6226 0.6133 0.6161 0.6102(0.1803) (0.1604) (0.3104) (0.0228) (0.0212) (0.0236) (0.0253)
λ2 0.4595 0.4991 0.5385 0.4132 0.4127 0.4135 0.4135(0.0620) (0.1144) (0.1623) (0.0163) (0.0164) (0.0147) (0.0154)
Bayes α 2.1554 2.1261 2.1173 2.0937 2.0718 2.0847 2.0601Prior 3 (0.1905) (0.1774) (0.1508) (0.1051) (0.0771) (0.0976) (0.0799)
λ1 0.6420 0.6541 0.6789 0.6232 0.6166 0.6139 0.6113(0.0557) (0.0484) (0.0571) (0.0221) (0.0207) (0.0193) (0.0222)
λ2 0.4329 0.4358 0.4475 0.4103 0.4120 0.4112 0.4177(0.0364) (0.0335) (0.0369) (0.0141) (0.0141) (0.0135) (0.0142)
Bayes α 2.4204 2.4818 2.5274 2.1774 2.1108 2.1713 2.1427Prior 4 (0.4894) (0.5103) (0.5547) (0.1366) (0.0963) (0.1285) (0.1051)
λ1 0.7464 0.8423 0.9306 0.6617 0.6554 0.6499 0.6532(0.1168) (0.1760) (0.2583) (0.0309) (0.0286) (0.0240) (0.6532)
λ2 0.4609 0.5220 0.5902 0.4331 0.4208 0.4262 0.4187(0.0518) (0.0727) (0.1174) (0.0166) (0.0155) (0.0168) (0.0158)
30
Table 5: The average confidence length/HPD credible length and coverage percentage inparentheses when α = 2, λ1 = 0.6 and λ2 = 0.4
Estimator CS-1 CS-2 CS-3 CS-4 CS-5 CS-6 CS-7MLE α 2.2293 2.1383 2.1969 1.3263 1.1607 1.2536 1.1465
(0.96) (0.96) (0.96) (0.95) (0.96) (0.96) (0.96)λ1 0.9906 1.1384 1.4709 0.5832 0.5900 0.5769 0.5856
(0.94) (0.95) (0.96) (0.94) (0.94) (0.94) (0.94)λ2 0.8038 0.8866 1.1094 0.4783 0.4772 0.4730 0.4756
(0.92) (0.95) (0.95) (0.93) (0.93) (0.93) (0.93)Bayes α 2.1871 2.0830 2.1110 1.3098 1.1272 1.2359 1.1139Prior 0 (0.95) (0.95) (0.95) (0.96) (0.94) (0.95) (0.94)
λ1 0.9763 1.0818 1.3769 0.5668 0.5688 0.5623 0.5568(0.93) (0.95) (0.97) (0.95) (0.95) (0.94) (0.93)
λ2 0.7649 0.8371 1.0087 0.4546 0.4572 0.4539 0.4493(0.93) (0.93) (0.94) (0.94) (0.94) (0.94) (0.94)
Bayes α 1.8102 1.6936 1.6673 1.2236 1.0694 1.1673 1.0512Prior 3 (0.98) (0.96) (0.97) (0.95) (0.96) (0.96) (0.95)
λ1 0.8520 0.8365 0.9267 0.5517 0.5545 0.5449 0.5518(0.95) (0.96) (0.97) (0.95) (0.95) (0.95) (0.94)
λ2 0.6805 0.6555 0.7075 0.4421 0.4464 0.4411 0.4476(0.93) (0.94) (0.95) (0.94) (0.94) (0.95) (0.94)
Bayes α 1.9970 1.9603 1.9687 1.2617 1.0749 1.2032 1.0785Prior 4 (0.93) (0.90) (0.89) (0.94) (0.94) (0.95) (0.94)
λ1 0.9589 1.0894 1.3283 0.5797 0.5807 0.5688 0.5695(0.94) (0.95) (0.96) (0.94) (0.94) (0.96) (0.95)
λ2 0.7244 0.8037 0.9694 0.4623 0.4562 0.4546 0.4493(0.92) (0.95) (0.97) (0.94) (0.95) (0.94) (0.94)
31
Table 6: The optimal censoring scheme for different objective functions when m = 5 and n= 10, 15, 20, 25 and 30. The relative efficiency (RE) and the relative expected time (RT) ofType-II censoring scheme with respect to the optimum censoring scheme are reported.
OF n R1 R2 R3 R4 R5 RE RT
1 10 1 4 0 0 0 96.1% 53.6%15 1 9 0 0 0 96.8% 46.0%20 0 15 0 0 0 98.7% 35.1%25 0 20 0 0 0 98.9% 33.5%30 0 0 25 0 0 99.5% 32.8%
2 10 4 1 0 0 0 92.9% 52.1%15 1 9 0 0 0 95.1% 46.0%20 0 15 0 0 0 96.8% 35.1%25 0 0 20 0 0 98.1% 37.8%30 0 0 25 0 0 98.8% 34.5%
3 10 0 0 0 0 5 100.0% 100.0%15 0 0 0 0 10 100.0% 100.0%20 0 0 0 0 15 100.0% 100.0%25 0 0 0 0 20 100.0% 100.0%30 0 0 0 1 24 99.9% 55.4%
4 10 5 0 0 0 0 89.4% 54.9%15 10 0 0 0 0 90.0% 43.0%20 15 0 0 0 0 91.3% 36.7%25 20 0 0 0 0 92.5% 32.5%30 25 0 0 0 0 93.4% 29.5%
5 10 5 0 0 0 0 98.0% 54.9%15 10 0 0 0 0 98.0% 43.0%20 15 0 0 0 0 98.4% 36.7%25 20 0 0 0 0 98.7% 32.5%30 25 0 0 0 0 98.9% 29.5%
32
Table 7: The optimal censoring scheme for different objective functions when n = 15, and m= 6, 8 and 10. The relative efficiency (RE) and the relative expected time (RT) of Type-IIcensoring scheme with respect to the optimum censoring scheme are reported.
m = 6
OF R1 R2 R3 R4 R5 R6 RE RT
1 1 8 0 0 0 0 96.8% 46.8%2 1 8 0 0 0 0 95.7% 46.8%3 0 0 0 0 0 9 100.0% 100.0%4 9 0 0 0 0 0 90.0% 46.1%5 9 0 0 0 0 0 98.2% 46.1%
m = 8
OF R1 R2 R3 R4 R5 R6 R7 R8 RE RT
1 1 6 0 0 0 0 0 0 98.6% 53.5%2 1 6 0 0 0 0 0 0 93.9% 53.5%3 0 0 0 0 0 0 0 7 100.0% 100.0%4 7 0 0 0 0 0 0 0 92.1% 52.4%5 0 7 0 0 0 0 0 0 97.8% 53.6%
m = 10
OF R1 R2 R3 R4 R5 R6 R7 R8 R9 R10 RE RT
1 0 1 4 0 0 0 0 0 0 0 97.9% 62.2%2 0 5 0 0 0 0 0 0 0 0 94.9% 61.7%3 0 0 0 0 0 0 0 0 0 5 100.0% 100.0%4 5 0 0 0 0 0 0 0 0 0 95.1% 61.2%5 0 2 3 0 0 0 0 0 0 0 98.4% 62.0%